Apply document groups through shared indexes

Groups are views over one complete example table. Build their indexes once and use the same indexes for every aligned field.

Goal

Apply the supplied document groups to examples without independent sorting, copying, or reordering of contexts, masks, targets, or metadata.

Inputs

Use the complete ordered table and the validated document_groups.csv. A document belongs to exactly one configured group: working, checking, or reserved. Every example inherits its document's group. The assignment is supplied; do not infer or randomize it.

Deliverables

Implement group selection in src/groups.py. Create one shared integer index array per group, then use it to select every aligned field. Save per-group metadata and arrays, output/group_summary.csv, group order, and the source assignment identity.

Checks

Validate that group assignments cover every input document once, name no unknown document, and use only configured names. Check that group example sets are disjoint and their union is the complete example set. Preserve the original order within each group.

Deliberately reorder target_ids or metadata independently and ensure the alignment check fails. Follow one composite identity from the complete table into its group arrays and confirm its context, mask, target, and metadata still describe the same row. Check a group with no documents and a group with no examples.

Workspace

Keep assignment validation in src/groups.py. Make one shared index the only selection boundary for aligned arrays. Do not batch or shuffle groups here.

Hints

Review

Pick one row from each non-empty group and trace its document and target position back to the complete table. Explain why no row can appear in two groups.

How to check your work

Checks compare group order, counts, disjointness, and identity preservation with the supplied fixture. The supplied fixture does not call these operating labels train, validation, or test sets.